Year: 2026
Type: conference-paper
Abstract: Complex reasoning in tool-augmented agent frameworks is inherently long-horizon, causing reasoning traces and transient tool artifacts to accumulate and strain the bounded working context of large language models.Without explicit memory mechanisms, such accumulation disrupts logical continuity and undermines task alignment.This positions memory not as an auxiliary efficiency concern, but as a core component for sustaining coherent, goal-directed reasoning over long horizons.We propose MemoBrain, an executive memory model for tool-augmented agents that constructs a dependency-aware memory over reasoning steps, capturing salient intermediate states and their logical relations.Operating as a co-pilot alongside the reasoning agent, Mem-oBrain organizes reasoning progress without blocking execution and actively manages the working context.Specifically, it prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget.Together, these mechanisms enable explicit cognitive control over reasoning trajectories rather than passive context accumulation.We evaluate MemoBrain on challenging longhorizon benchmarks, including GAIA, Web-Walker, and BrowseComp-Plus, demonstrating consistent improvements over strong baselines.Our codes are in this repository.Episode 1 Main Agent Episode 2 Episode 3 … Episode t MemoBrain Episode t (Reasoning Trace) Think Hmm… these results seem noisy.Not sure if they help… Maybe try refining the query… Tool Result search(query="Q refined", k=10) search(query="Q alternative", k=10) … 10 pages returned, mixed relevance.Transient, noisy execution traces during ongoing trial-and-error.up to thousands of tokens Thought 1 Thought 2 Thought 3 … Thought t Thought t (Memory Unit) Built upon prior unresolved search (Thought 7) Refine search strategy to resolve missing evidence Relevant evidence identified for subproblem Q Resolved reasoning decision abstracted as a persistent memory unit.typically ~10× more compact Context Decision Outcome Reasoning Context Budget Reached Asynchronously Memorize 1.
Authors: Hongjin Qian, Zhao Cao, Zheng Liu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166803090